I just ran Qwen3.5 35B on my iPhone at 5.6 tok/sec.
On 22 March, a user on LocalLLaMA posted that they had run Qwen3.5 35B on an iPhone, hitting 5.6 tokens per second. That is the complete evidentiary record. One post, one person, one number — and nothing has moved since. The thread scored 8.2 in LocalLLaMA's community ranking, which tells you the claim caught people's attention, not that anyone reproduced it. No screen recording surfaced in the thread, no second user posted matching results, no follow-up from the original poster elaborating on their setup. The story is exactly where it started a month ago: interesting, unverified, and waiting.
If confirmed, here is what this means. A 35-billion-parameter model running at conversational speed on consumer mobile hardware is not a benchmark curiosity — it is a quiet dismantling of the assumption that serious AI inference requires a data centre, or even a desktop. Five tokens per second is slow enough to notice in a chat interface but fast enough to use. That threshold matters. It means a model with genuine reasoning depth sitting entirely on-device, no API call, no subscription, no connectivity required. The privacy implications alone reshape how people think about deploying LLMs in sensitive contexts — healthcare notes, legal drafts, personal journaling. For Apple, it raises questions about whether their own on-device model strategy is moving fast enough. For Qualcomm and the broader edge-inference ecosystem, it is validation that the hardware investment is beginning to compound. The second-order effect is slower but larger: if this is reproducible on a current iPhone, it becomes reproducible on last year's phone within eighteen months, and the installed base of capable inference devices grows by hundreds of millions without anyone buying anything new.
Watch for a second user posting reproducible results with visible proof — hardware model, quantisation method, and a screen recording. If that appears, this moves from background noise to a story worth tracking seriously.
<table> <tr><td> <a href="https://www.reddit.com/r/LocalLLaMA/comments/1s3i83m/fully_local_voice_ai_on_iphone/"> <img alt="Fully local voice AI on iPhone"
<!-- SC_OFF --><div class="md"><p>Long post, but hopefully helps somebody. Llama-cpp vulkan server running single AMD R9700. The settings below are showing great results with a
<!-- SC_OFF --><div class="md"><p>I can't believe it, but I'm able to do my daily software development work on this model. We have a 500-700k line of code enterprise software
<table> <tr><td> <a href="https://www.reddit.com/r/singularity/comments/1s7rcwr/qwen35_omni_qwens_latest_generation_of_fully/"> <img alt="Qwen3.5 Omni - Qwen’s latest generation
<!-- SC_OFF --><div class="md"><p>I'm genuinely unsure if I'm doing something wrong, or if the model is just worse, despite the community enthusiasm. I tend to use it with pi
<!-- SC_OFF --><div class="md"><p>I’ve been experimenting with running LLMs fully on-device, and managed to get Gemma 4 running locally on an iPhone 13 Pro.</p> <p>This is built
<!-- SC_OFF --><div class="md"><p>I could not find good data points on what speed one could get with a single 5090 and enough DDR4 RAM. </p> <p>My system: AMD EPYC 7532 32core
<!-- SC_OFF --><div class="md"><p>I'm getting ~13 tps on Q8_0, with a context window of 128000, K Q8_0, V Q8_0</p> <p>this is on 3x GPUS (1x2060super 8gb, 2x5060ti 16gb), via
NewsHive monitors these sources continuously. All signal titles above link to the original reporting.
Intelligence by NewsHive. Need help navigating what this means for your business? Contact GeekyBee →